LoCon and LoHA each fail to fully replace LoRA, for opposite reasons: LoCon is too broad, LoHA is too narrow.However, when this LoCon is stacked at very low weight (0.1–0.2) with my other LoRA <a target="_blank" rel="ugc" href="https://civarchive.com/models/2369235/ghost-and-darkness-fate-lancer-x-archer-role-fixed-no-swap">([link])</a>, it significantly improves spatial composition and canvas utilization beyond what the LoRA achieves alone. For details, refer to my article (<a target="_blank" rel="ugc" href="https://civarchive.com/articles/26891/why-you-still-need-locon">[CivitAI article link])</a>.This model was trained on a Pony base, but as discussed in the article, it is also effective on Illustrious-based models. You may need to lower the weight further depending on your specific model. Don't worry about going too low—this LoCon performs its role fully even at 0.1. Solo use is possible but not recommended.<h3 id="faq">FAQ</h3>Q. Why release a LoCon in an overfitted state?While other aspects are overfitted, the key differentiator of LoCon over LoRA—spatial awareness via Conv layers—performs exactly as intended. Stacking at low weight is essentially cherry-picking only that Conv-layer spatial perception.Q. Why not merge it with other LoRAs, or bake into a checkpoint and re-extract?Conventional wisdom says LoRA merging is nearly lossless, and bake-then-extract is considered the most advanced technique. I have naturally tried all of these approaches, including retraining. However, the relational dynamics of a male coupling—specifically, the balance of who leads and who follows—proved far too delicate to survive any of those processes.Q. Do I really need to stack multiple LoRAs just for this?Personally, I use a merged version combining my published LoRA with this LoCon, and I'm quite satisfied with it for private use. However, it's currently difficult to control, with minor issues like unintended particle effects appearing, so I won't be releasing that merged version.<h3 id="notes">Notes</h3><ol><li>The dataset was built from my own illustrations, processed through hires fix and img2img.</li><li>Merging with other LoRAs, LoCons, or checkpoints is permitted, but redistribution of any merged result—including the original—is prohibited.</li><li>While I do not assert full copyright over the dataset, this model is ultimately based on my artwork, and commercial use by others is not permitted.</li></ol>

GND - Spatial Awareness Enhancement LoCon

It's based on pony but will works on illustrious.

Files with SHA256: 461c532c5f761b75e840d04ca2efc72d6c412f3e199cd32fd4cc249b1376df75